This work presents a hybrid wrapper/filter algorithm for feature subset selection that can use a combination of several quality criteria measures to rank the set of features of a dataset. These ranked features are used to prune the search space of subsets of possible features such that the number of times the wrapper executes the learning algorithm for a dataset with M features is reduced to O(M) runs. Experimental results using 14 datasets show that, for most of the datasets, the AUC assessed using the reduced feature set is comparable to the AUC of the model constructed using all the features. Furthermore, the algorithm archieved a good reduction in the number of features.Sociedad Argentina de Informática e Investigación Operativa